Focal vs Diffuse: Mechanisms of attention mediated performance enhancement in a hierarchical model of the visual system

Abstract

Spatial attention is an essential cognitive process for visual perception, especially in complex scenes with poor luminance. Attentional modulation of neural activity has been documented across the visual cortex. However, how these changes in neural response lead to better behavioral performance on challenging tasks remain unknown. In our study, we implemented spatial attention in a deep convolutional neural network model of the ventral visual hierarchy and measured its impact on categorization performance on cluttered images with varying contrast levels. We applied attention to network units in three ways: enhancement in attended region (EAR), suppression in unattended region (SUAR) or both. When focally applied to a single convolutional layer, SUAR is more effective in boosting performance than EAR, especially in the presence of high contrast distractors. SUAR is also effective in recovering degraded performance due to low contrast targets, whereas EAR fails to recover. These results predict a novel mechanism of suppression of neural activity corresponding to the unattended parts of visual space. Intriguingly, EAR in our model achieves the same performance as SUAR when attention is diffusely applied to stacks of successive convolutional layers, irrespective of target contrast. This suggests an alternate mechanism of attention wherein enhancement of attended neural activity alone, when applied to successive cortical encoding stages, is an effective strategy for boosting performance in challenging object recognition tasks. Our results indicate that the two alternative attentional mechanisms are functionally equivalent in tackling challenging object recognition tasks.